The nodes in a network can be grouped into 'roles' based on similar connection patterns. This is usually achieved by defining a pairwise node similarity matrix and then clustering rows and columns of this matrix. This paper presents a new similarity matrix for solving role extraction problems in directed networks, which is defined as the solution of a matrix equation and computes node similarities based on random walks that can proceed along the link direction and in the opposite direction. The resulting node similarity measure performs remarkably in role extraction tasks on directed networks with heterogeneous node degree distributions.